For more information,

ebusiness@mit.edu or 617-253-7054
please visit our website at http://ebusiness.mit.edu
or contact the Center directly at
A research and education initiative at the MIT
Sloan School of Management
Does Current Advertising Cause Future Sales?
Evidence from the Direct Mail Industry

Paper 222





Duncan Simester
Yu (Jeffrey) Hu

Erik Brynjolfsson

Eric T. Anderson


January 2006



Does Current Advertising Cause Future Sales?
Evidence from the Direct Mail Industry

December, 2005

Duncan Simester
MIT Sloan School of Management
simester@mit.edu

Yu (Jeffrey) Hu
Purdue University
yuhu@purdue.edu


Erik Brynjolfsson
MIT Sloan School of Management
erikb@mit.edu

Eric T. Anderson
Northwestern University
eric-anderson@kellogg.northwestern.edu


Abstract
We present findings from a large-scale field experiment that allows us to study whether
there is a causal relationship between current advertising and future sales. The
experimental design overcomes limitations that have affected previous investigations of
this issue. We find that current advertising does affect future sales but the sign of the
effect varies depending on the customers targeted. For the firms best customers the
long-run effect of increases in current advertising is actually negative, while for other
customers the effect is positive. We argue that these outcomes reflect two competing
effects: brand-switching and inter-temporal substitution. Furthermore, our data suggest a
way to distinguish between the informative and persuasive roles of advertising, providing
insight into the mechanism by which advertising differentially affects various customer
subsets.
Acknowledgements: We thank seminar participants at Georgia Institute of Technology,
MIT, Northwestern University, Purdue University, University of Connecticut, University
of Maryland, University of Pennsylvania, University of Southern California, 2004
Workshop on Information System and Economics, 2005 Symposium on Electronic Commerce
Research, Fifth Annual INFORMS Revenue Management and Pricing Conference. Generous
funding was provided by MIT Center for eBusiness.
Does Current Advertising Cause Future Sales? Page 1
1. Introduction
Over $245 billion was spent on advertising in the United States in 2003 (Advertising
Age, 2005). Despite enormous levels of spending, an important economic question has
yet to be resolved. Is there a positive association between current advertising and future
sales? Years of academic research investigating the relationship between advertising and
future demand has yielded inconclusive findings. Most of these previous studies have
been plagued by two obstacles:
1. Advertising decisions are endogenous and so effects attributed to variations in
advertising expenditure may actually reflect factors that led to the variation in
expenditure.
2. Advertising is dynamic and so the effects of advertising are often confounded by
other intervening effects.

These obstacles are well-recognized and a variety of econometric methods have been
proposed to address them. In this paper we report findings from a largescale field test in
which we address both problems using an alternative approach. We experimentally vary
advertising strategies for two randomly selected Treatment and Control samples of
customers. This experimental manipulation introduces an exogenous source of variation
that overcomes the endogeneity limitation that has limited previous studies. Both
samples are exposed to the same intervening events such as competitive reactions or
macro-economic changes. Thus, the comparison between the Treatment and Control
samples overcomes these potential confounds.

The field test was conducted in the direct mail industry with a mail-order catalog that
sells womens clothing in the moderate price range. Direct mail represents 20% of all
advertising in the United States and is the single largest type of media advertising
(Advertising Age, 2005). As a comparison, 2003 advertising spending was $48 billion in
direct mail, $45 billion in newspapers and $42 billion in broadcast television. In the
United States direct mail companies mailed over 18 billion catalogs in 2004 (Direct
Marketing Association, 2005). The catalogs typically announce a companys retail prices
and contain detailed information about the available products, together with information
about ordering procedures, warranties, and payment methods. This focus on providing
Does Current Advertising Cause Future Sales? Page 2
information about available products and prices is typical of advertising by retailers (and
contrasts with advertising by manufacturers).

It is well accepted in the direct mail industry that catalog advertising increases current
demand. What is much less well-understood is the impact that it has on demand in future
periods. With few exceptions, companies adopt a myopic focus when optimizing their
catalog advertising decisions. They estimate the probability that a customer will order
from a specific catalog and then mail to all customers for whom the expected lift in
immediate demand justifies the incremental printing and mailing costs. This focus solely
on the current impact of catalog advertising either implicitly presumes that there is no
long-term effect or simply ignores any long-term effects.

The possibility that catalog advertising may have a long-run impact on demand has been
recognized in the academic literature. Indeed, models have been proposed to help
companies solve the difficult dynamic optimization problem that arises if catalogs do
impact future demand (see for example Gnl and Shi, 1998; and Simester, Sun and
Tsitsiklis 2005). However, there are apparently no published studies directly estimating
the impact of catalog advertising on future demand or comparing how these long-run
effects (if any) vary across customers.

1.1 Prior Theoretical Work
Does current advertising increase future sales? Unfortunately, existing theory and
evidence provides two conflicting answers to this basic question. Much of the theoretical
advertising literature has focused on distinguishing whether advertising serves a
persuasive or informative role. Under the persuasive view, advertising alters
customers utility functions by changing their preferences (Kaldor 1950; Comanor and
Wilson 1967 and 1974; and Becker and Murphy 1993). This leads to an outward shift in
the demand function, which has led to claims that advertising may serve an important
anti-competitive role. Under the informative view, advertising does not change
customers utility functions (Stigler 1961; Kihlstrom and Riordan 1984; and Milgrom and
Does Current Advertising Cause Future Sales? Page 3
Roberts 1986).
1
Instead it increases the information that customers have about the
available alternatives.

Both views of the role of advertising predict an increase in customers expected utility
from consuming the advertised product. Under the persuasive view the change in
expected utility reflects a change in the utility function itself, while the informative view
predicts a change in which alternatives are evaluated and/or what is known about those
alternatives. Both outcomes are consistent with advertising positively impacting demand
in future periods. In particular, under the persuasive view we would generally expect a
change in the utility function to endure into future periods. Similarly, product
information revealed under the informative view will generally be relevant in future
periods, unless changes between periods make past information obsolete.

Yet it is also possible that the long-run impact of advertising is negative. When making
purchasing decisions customers generally have the alternatives of purchasing competing
brands, purchasing from different retailers or even delaying in the hope of future
discounts or product improvements. If advertising makes an immediate purchase of the
focal brand more attractive, it implicitly reduces the share of customers who will choose
one of these alternatives. The outcome is potentially less demand for competing brands,
less demand for competing retailers and/or less demand in future periods. Of these
outcomes, the impact on competing brands (sometimes termed the combative role of
advertising) has received the most interest. As early as 1942 Borden distinguished
between the primary and selective effects of advertising: the primary effect describes
category-level demand expansion, while the selective effect describes substitution
between competing brands. More recently, the distinction between advertisings primary
and selective effects has served as a central focus of debate in the tobacco industry (see
for example Seldon and Doroodian 1989; and Roberts and Samuelson 1988). The
industry has sought to ward off proposed regulation limiting tobacco advertising by
arguing that advertising serves primarily a selective role, allowing companies to attract
share from their competitors without expanding total industry demand. In contrast, anti-

1
See also: Telser 1964; Nelson 1970 and 1974; Schmalensee 1978; and Grossman and Shapiro 1984.
Does Current Advertising Cause Future Sales? Page 4
smoking advocates have argued that tobacco advertising also has an impact on primary
demand, contributing to an expansion in total tobacco consumption.

Substitution between brands is analogous to substitution across time. In many product
categories purchasing a competing brand and purchasing in future periods both represent
alternatives to making an immediate purchase of the focal brand. Although the
possibility of inter-temporal substitution has received relatively little attention in the
advertising literature, it has received considerable attention in the pricing literature.
There is well-documented evidence that price discounts can lead to both brand-
substitution and inter-temporal substitution. As a result, following a price promotion,
there is often evidence of a post-promotion dip in sales, as customers consume products
purchased during the discount period (Blattberg and Neslin 1990, p. 358; and Hendel and
Nevo 2003).
2
Interestingly, there is also evidence that this inter-temporal effect varies
across customers (Anderson and Simester 2004). The negative long-run effect of a price
promotion appears to be most pronounced for customers who have the most experience
with the brand.

We conclude that there is theoretical support for advertising having both a positive and a
negative impact on future demand. If advertising increases customers expected utility
through persuasion or information, and this increase is enduring, the impact on future
demand will tend to be positive. On the other hand, if advertising accelerates demand,
temporal-substitution may lead to a negative impact on future demand.
1.2

Prior Empirical Evidence
The previous empirical evidence is mixed. There is some evidence of a positive long-run
relationship between advertising and sales. Yet many studies report either no long-run
impact or that the impact is short-lived (Bagwell 2005). There have apparently not been
any studies reporting a negative relationship between advertising and future demand.
However, as we recognized, this empirical work has been confronted by important
challenges. The earlier work was typically limited to aggregate brand or category-level

2
See also Hendel and Nevo (2002 and 2005).
Does Current Advertising Cause Future Sales? Page 5
data, in which researchers investigated the relationship between current advertising and
lagged effects on sales. Because the sign of the effect could theoretically vary for
different subsets of consumers, aggregate data may not detect this relationship even when
it is present. These studies also suffered from important limitations due to the
endogeneity of the advertising decisions, since changes in sales can lead to changes in
advertising budgets, and confounds introduced by intervening events (Schmalensee 1972;
and Lambin 1976). More recently, the development of household level panel datasets has
made it possible to estimate demand at the individual or household-level. Together with
methodological developments in the estimation of simultaneous structural models, these
new datasets offer the opportunity to address endogeneity through advanced econometric
controls (see for example: Erdem and Keane 1996; and Ackerberg 2003).

In contrast, our approach has not been to exploit ever more sophisticated econometric
methods but rather to improve the direct measures of advertising and sales. In particular,
the direct sales industry provides a particularly measurable domain for studying
responses to advertising. Furthermore, the experimental approach that we adopt in this
paper departs from most earlier attempts to overcome endogeneity and intervening effects
by introducing random assignment of customers to Treatment and Control groups
with external controls to the data collection process to prevent the introduction of
confounds. This contrasts with previous studies in which researchers have had to accept
the presence of confounds in their data and instead sought to provide internal controls for
these confounds in their analyses. The experimental approach also offers another
advantage: the results are easily analyzed and interpreted. The experimental design
yields a simple comparison between groups of customers who experience one advertising
treatment and equivalent control groups who experience a different treatment. We
directly measure the difference in their long-run demand. The outcome is immediately
interpretable.

This is not the first field experiment designed to investigate the impact of advertising.
Managerial studies using proprietary split-sample cable TV experiments have previously
been used in the consumer packaged goods industry. Unfortunately, academic
Does Current Advertising Cause Future Sales? Page 6
descriptions of these findings are necessarily limited by the proprietary nature of the data
and estimation models (see for example: Aaker and Carmen 1982; Lodish et al. 1995a
and 1995b). Moreover, the results are apparently mixed, perhaps in part due to a lack of
statistical power.
3
In addition, there have been at least two academic studies that use
experiments to investigate how advertising influences prices and price elasticities.
Krishnamurthi and Raj (1985) also report the findings from a split sample cable TV
experiment, and conclude that advertising is capable of reducing consumer price
elasticities. More recently Milyo and Waldfogel (1999) use a natural experiment to study
the effect of advertising on prices. They find that advertising does tend to lower the prices
of advertised products, but has little effect on the prices of unadvertised products.
4

1.3

Preview of the Key Findings
The field experiment reported in our paper reveals several findings. First, as we would
expect, customers in the Treatment condition who received additional catalogs purchased
more items in the short-term than their counterparts in the Control condition. This result
confirms that current advertising can increase current demand. As we discussed, this
finding is already well-understood in the direct mail industry but is also the limit of most
firms analyses. These firms do not look beyond the immediate impact to also consider
how catalog advertising affects demand in the future or demand in other channels.

Our findings confirm that current advertising also has a significant impact on demand in
future periods. However, the effect is not always positive. Among the catalogs Best
customers, who had historically purchased recently and frequently, increased current
advertising significantly reduced future demand. We interpret this result as evidence of
temporal substitution. In fact, we find that the short-run increase in demand among these
customers is almost entirely offset by the reduction in future demand. Further
investigation also revealed evidence of cross-channel substitution: the increase in demand

3
These three papers do not report sample sizes or the estimation models for individual studies.
4
For other examples of natural experiments see: Benham (1972) and Ippolito and Mathios (1990).
Does Current Advertising Cause Future Sales? Page 7
from the catalog channel (mail and telephone orders) was offset by a reduction in demand
from the Internet channel.

In contrast, for the Other customers in the study, who had historically purchased less
frequently and/or less recently than the Best customers, the results were reversed.
Sending more current catalogs led to an increase in future demand.

Why do companies mail so many catalogs to their best customers? Our findings may
provide an explanation. It is not uncommon for companies to send catalogs every two
weeks, with some companies sending their best customers as many as 100 catalogs a
year. These intensive mailing policies often prompt complaints from customers that they
receive too many catalogs. Yet paper and postage are not free and so these policies are
not arbitrary. One explanation is that firms are myopic and limit their attention to the
short-run impact of mailing decisions on purchases from the catalog channel. If
companies were to extend their analysis to also consider the long-run and cross-channel
impacts, they would learn that for their best customers, the lift in short-run demand is
largely due to inter-temporal and cross-channel substitution. In later discussion we offer
an explanation for why companies commonly overlook these externalities.

1.4 Structure of the Paper
The paper proceeds in Section 2 with a simple model illustrating the intuition that current
advertising may lead to a positive or a negative impact on future demand. We then
provide an overview of the study design in Section 3 before presenting the results in
Section 4. The results section begins with a review of the short-run impact followed by
the long-run and cross-channel outcomes. We then investigate alternative explanations
for the findings by comparing the heterogeneity in the results across different customer
segments. The paper concludes in Section 5 with a review of the findings and
implications.

Does Current Advertising Cause Future Sales? Page 8
2. Positive and Negative Long-Run Outcomes
To help understand why current advertising may lead to a positive or a negative impact
on future demand we present a stylized model that highlights two opposing advertising
outcomes: brand-switching and inter-temporal substitution.

We consider a two period problem in which a firm produces different products each
period. In the first period customers decide how many products to purchase and consume
in that period ( ) and how many products to purchase and stockpile for the next period
( ). In the second period, customers choose how many additional units to purchase and
consume in that period ( ). Stockpiling between periods allows customers to introduce
variety to their second period consumption decisions (recall that the firm produces
different products each period). For simplicity we assume that the price charged by the
focal firm (p) does not vary between periods and set the inter-period discount rate to 1.
These assumptions jointly ensure that in this model stockpiling cannot be explained by
mere price arbitrage.
1
q
1
i
2
q

Customers can also choose to purchase a competing brand. We use an upperbar to
distinguish the price of the competing brand ( p ) and the customers quantity decisions
(
t
q ) for this competitor. We set the competitors price to one and do not consider stock-
piling of this outside option as it adds few additional insights and does not alter our key
findings. Finally, we also assume that there is a budget constraint such that:
( ) ( )
1 1 2 1 2
Y p q i q p q q = + + + + .

To simplify the analysis and exposition it is helpful to describe consumption utility using
a separable quadratic function:

( ) ( ) ( )
1 1 1 1 1 1 1
, U q q q v q q v q = + (2)
( ) ( ) ( ) ( )
1 2 2 1 1 1 2 2 2 2 2
, , U i q q i v i q v q q v q = + + (3)

Does Current Advertising Cause Future Sales? Page 9
The terms are preference parameters that are influenced by advertising.
t
v
5
We make the
natural assumptions that is increasing in both current and prior period advertising, and
that carryover to future periods decays over time: , while
. We also assume that advertising by the focal firm does not
directly affect preferences for the competing product:
t
v
/ / 0
t t t t j
dv da dv da j

> > 0 >

0 / 0
t t k
dv da k
+
= >
/
t
dv da 0 = . While the
relationship between advertising and preferences for the focal brand is positive, this
illustrative model does not speak to the source of this relationship. In particular, we do
not seek to distinguish between the information and persuasion interpretations proposed
in the literature. We later use our empirical findings to investigate this issue.

Customers select the quantity of goods that maximizes utility for both periods, subject to
their budget constraint. Solving the resulting system of first-order conditions reveals
customers optimal consumption decisions:
* * 1 2
1 1
3 2(
10
v v Y v
q i
+
= =
)
(4)
* 2 1
2
2 (
5
v v Y v
q
+
=
)
(5)
* * 1 2
1 2
3 2 (2
10
v Y v v
q q
+ +
= =
)
(6)
The key insights concern the relationship between advertising in period 1 and customers
purchasing decisions of the inside goods.
* *
1 1 1 2
1 1 1 1
1
3
3
dq di dv dv
da da da da

= = >

(7)
*
2 2
1 1
1
2
3
dq dv dv
da da da

=

1
1

0

(8)
As we would expect, the impact of period 1 advertising on period 1 demand is positive:
and . The impact on future demand ( ) is ambiguous and
*
1 1
/ 0 dq da >
*
1 1
/ di da >
*
2
q

5
We assume that , which ensures that t v Y
t
2 t v q
t t
2 and therefore consumption utility is
always increasing in quantity. In practice, as long as customers will always prefer to choose and
less than . We also make analogous assumptions for the competitive product.
0 p >
t
q
t
i
t
v
Does Current Advertising Cause Future Sales? Page 10
reflects a trade-off between brand-switching and inter-temporal substitution. In
particular, the sign of depends upon the rate at which the long-run impact of
advertising decays. Because advertising in period 1 has a favorable impact on
preferences in period 2 ( ) it leads to a switch in second period demand from
the competing brand to the focal brand. Yet current advertising has a bigger impact on
current preferences than future preferences ( ), and so the lift in second
period demand is offset by forward-buying in which customers shift second period
demand to the first period.
*
2 1
/ dq da > 0
0
1
/
2 1
/ dv da >
2 1 1
/ dv da dv da <

This analysis also suggests that the relationship between current advertising and future
demand will vary across customer segments. If customers already have strong
preferences for the brand, there is relatively little opportunity for additional brand-
switching. For example, consider a segment of customers whose preferences for the
focal firm are so strong that they do not purchase any units from the competing brand.
After setting
* *
1 2
0 q q = = and maximizing utility subject to ( )
1 1 2
Y q i q = + + , the first
order condition for yields the following second period demand:
2
q
* 2 1
2
3
v v Y
q
+
= (9)
Among consumers who never purchase the outside goods, the long-run impact of
advertising is no longer ambiguous:
*
2 1
/ dq da 0 . Sending additional advertising to these
customers cannot lead to any further brand-switching, and so the only remaining effect is
inter-temporal substitution. We conclude that for customers with very strong preferences
for the firm, current advertising may lead to a reduction in future demand. In contrast,
among customers with weaker ex ante preferences for the firm, the possibility of brand
switching is more likely to lead to a favorable long-run outcome.

The study described in the next section provides an opportunity to test these predictions.
We can use customers transaction histories (prior to the study) to distinguish customers
with strong preferences for the focal firm from those with weaker preferences. Random
assignment yields equivalent Treatment and Control samples of both types of customers.
Does Current Advertising Cause Future Sales? Page 11
Therefore, comparing how the response to the experimental manipulation varied across
these samples reveals how prior preferences moderate the long-run impact of advertising.

3. Study Design
The study was conducted with a medium-sized company that sells womens clothing in
the moderate price range.
6
All of the products carry the companys private label brand
and are sold exclusively through the companys own catalogs, Internet website and retail
stores. The study involved a total of 20,000 customers who had previously made a mail
or telephone purchase from the company. To explore the effects of heterogeneity in the
sample, the company initially identified two distinct samples of customers. The first
sample of 10,000 customers, which we denote the Best customers, were all customers
who had made relatively frequent and recent purchases from the company. In particular,
these were the customers whom the companys own statistical models suggested would
be most likely to purchase if mailed a catalog.
7
The Other sample of 10,000 customers
was comprised of customers who the companys statistical model predicted had an
average probability of responding if mailed a catalog.

Random assignment was then used to assign these two samples of customers into equal
sized Treatment and Control groups. This yielded a total of four different customer
samples (see Table 1). In each case the final sample sizes were slightly smaller than
5,000. The reason for this is rather technical but does not affect the interpretation of the
study.
8



6
The company asked to remain anonymous.
7
Although the details of the companys statistical models are proprietary and were not made available to
the research team, the recency and frequency of prior purchases accurately distinguish these customers.
8
Because customers rarely have their unique customer identification numbers available when they call to
place an order, individual customers sometimes end up with more than one account number. Each month
the company uses various methods to identify these duplicate account numbers and consolidate them back
to a single account number. The reduction in the sample sizes reflects the deletion of duplicate account
numbers. Fortunately this process is identical for the Treatment and Control samples and so cannot explain
systematic differences between them.
Does Current Advertising Cause Future Sales? Page 12
Table 1: Sample Sizes

Control Sample Treatment Sample
Best Customers 4,921 4,904
Other Customers 4,790 4,758


The experimental manipulation occurred over an (approximately) eight-month period.
During this period all of the customers in the Treatment sample received a total of
seventeen catalogs, while customers in the Control sample received just twelve catalogs.
The additional catalogs sent to the Treatment sample were simply additional copies of
catalogs that all customers received. This ensured that the experimental manipulation
only affected the frequency of advertising, and not which products were available or
features specific to the design of the catalogs. Sending multiple copies of the same
catalog to the same customer is a common practice in the catalog industry. The cost of
designing new catalogs is expensive and so rather than designing new catalogs companies
will often re-send the same catalog two to four weeks after the first mailing.

The actual mailing schedule for the two samples is summarized in Table 2. The specific
timing of each mailing was determined by the companys circulation managers. The
managers were instructed to optimize the overall (short-run) response given the
exogenous decision to mail a total of twelve times to the Control sample and seventeen
times to the Treatment samples. It is possible that varying the timings would lead to
differences in the long-run results. Following the experimental manipulations the
company returned to using its standard circulation procedures to decide who to mail
catalogs to, and made no distinction between customers in the Treatment and Control
samples.


Does Current Advertising Cause Future Sales? Page 13
Table 2: Mailing Dates in 2002 by Experimental Condition
Control Treatment
Catalog 1
Mailing Date 1 January 11 January 11
Mailing Date 2 February 22 February 8
Catalog 2
Mailing Date 1 February 1 January 25
Mailing Date 2 February 22
Catalog 3
Mailing Date 1 March 15 March 8
Mailing Date 2 April 26 April 5
Catalog 4
Mailing Date 1 April 5 March 22
Mailing Date 2 May 3
Catalog 5
Mailing Date 1 May 17 April 19
Mailing Date 2 May 17
Catalog 6
Mailing Date 1 June 7 June 7
Mailing Date 2 June 28 June 28
Catalog 7
Mailing Date 1 July 26 July 26
Mailing Date 2 September 6 August 23
Mailing Date 3 September 20
Catalog 8
Mailing Date 1 August 9 August 9
Mailing Date 2 September 6

Because the first catalog was mailed to both samples on the same day, the date of the first
manipulation was actually January 25, 2002 (when only customers in the Treatment
group were sent Catalog 2). The last date on which the mailing dates were different for
the two samples was September 20, 2002. We received data describing the number of
items purchased by customers before, during, and after the experimental manipulations.
In particular, we received a record of all transactions made from January 1, 1988 until
almost nineteen months after the start of the first manipulation (August 13, 2003). To
simplify the analysis and discussion of the results, it is helpful to define three periods:

Does Current Advertising Cause Future Sales? Page 14
1. The Pretest period: from January 1, 1988 through January 24, 2002.
2. The Test period: from January 25, 2002 through December 31, 2002.
3. The Posttest period: from January 1, 2003 through August 13, 2003.

Notice that the Test period extends for 103 days beyond the date of the last manipulation:
September 20, 2002 through December 31, 2002. This was designed to capture orders
from catalogs mailed towards the end of the manipulation period. The company
estimated that over 99% of the immediate demand from catalogs mailed in September
would have occurred by December 31. This is also consistent with the industry-wide
response curve reported by the DMA (Direct Marketing Association 2003). We later
vary the length of the Posttest period to investigate how it affects the results (see Table
6).

We caution that the transaction data only involves customers purchases through the
companys Internet website or its catalog channel (mail and telephone orders). We do
not have a record of purchases made by these customers in the companys retail stores
because at the time of the study the company was unable to adequately identify customers
purchasing in its stores. We will later discuss how this omission may have affected the
results.

The historical purchasing results provide a means of checking whether the assignment of
customers to the Treatment and Control conditions was truly random. In particular, in
Table 3 we compare the average Recency, Frequency and Monetary Value (RFM) of
customers purchases during the Pretest period.
9
If the random assignment was truly
random we should not observe any systematic differences in these historical measures
between the Treatment and Control samples. The findings reveal no significant
differences in the historical demand in either the Best-customer or Other-customer
comparisons.

9
Recency is measured as the number of days (in hundreds) since a customers last purchase.
Frequency measures the number of items that customers previously purchased. Monetary Value
measures the average price (in dollars) of the items ordered by each customer.
Does Current Advertising Cause Future Sales? Page 15
Table 3: Check on Randomization Process
Purchases During the Pretest Period
Control
Condition
Treatment
Condition
p-value
Best Customers
Recency
1.43
(0.02)
1.43
(0.01)
0.72
Frequency
40.38
(0.45)
40.75
(0.51)
0.59
Monetary Value
61.11
(0.19)
61.22
(0.19)
0.69
Sample Size 4,921 4,904
Other Customers

Recency
4.67
(0.06)
4.76
(0.06)
0.30
Frequency
10.56
(0.20)
10.62
(0.21)
0.85
Monetary Value
63.85
(0.29)
64.18
(0.33)
0.50
Sample Size 4,790 4,758
The table reports the average values of each variable for each sub-sample.
Standard errors are in parentheses. The p-value denotes the probability that
the difference between the Treatment and Control averages will be larger
than the observed difference (under the null hypothesis that the true
averages are identical).

4. Results
4.1 Does Current Advertising Impact Short-Run Demand?
In Table 4 we summarize demand in the Treatment and Control conditions during the
Test period and report both univariate and multivariate comparisons. The univariate
analysis is simply the average number of items purchased by customers in each sample.
The multivariate analysis uses customers pretest purchases to control for individual
customer characteristics. In particular, the Recency, Frequency and Monetary Value
(RFM) of customers prior purchases, which we used to check the validity of the
randomization procedures (see Table 3), are well-established metrics for segmenting
customers in this industry and provide natural candidates for control variables. The unit
Does Current Advertising Cause Future Sales? Page 16
of analysis in the multivariate analysis is a customer (denoted by subscript i), and the
dependent measure is the number of items purchased during the Test period (Q
i
).
Because Q
i
is a count measure, the multivariate analysis uses Poisson regression. In
particular, we assume that Q
i
is drawn from a Poisson distribution with parameter
i
:

( ) Prob , =0, 1, 2, ...
!

= =
i
q
i
i
e
Q q q
q
(10)

where: ( ) ln
i
= X
i
i
. The X
i
terms denote the independent variables, which include the
log of each of the three RFM measures. To evaluate the impact of the experimental
manipulation we include a dummy variable identifying whether customer i was in the
Treatment condition. This yields the following model:

0 1 2 3
4
log( ) log( ) log( )
i i i
i
Recency Frequency Monetary Value
Treatment
= + + +
+
X
(11)

This analysis preserves the benefits of the experimental design. Under this specification,

4
measures the percentage change in short-run demand between customers in the
Treatment condition compared to those in the Control. This comparison with the Control
provides an explicit control for intervening factors, such as competitors actions and
macro-economic factors. We estimated separate models for the Best and Other
customers.

Does Current Advertising Cause Future Sales? Page 17
Table 4: Units Ordered During the Test Period

Other
Customers
Best
Customers
Univariate Analysis
Control Condition
1.08
(0.04)
3.63
(0.08)
Treatment Condition
1.24
(0.05)
3.86
(0.09)
Difference
0.16
*
(0.07)
0.23
*
(0.12)
Sample Size
9,548
9,825
Multivariate Analysis
Intercept
-1.213
**
(0.133)
-4.255
**
(0.117)
Recency
-0.276
**
(0.006)
-0.131
**
(0.004)
Frequency
0.489
**
(0.010)
0.749
**
(0.008)
Monetary Value
0.424
**
(0.029)
0.827
**
(0.026)
Treatment
0.138
**
(0.019)
0.051
**
(0.010)
Log Likelihood
-19,160
-33,919
Sample Size
9,548
9,825
The univariate analysis reports the average number of units
purchased during the Test period. The multivariate analysis
reports the coefficients from Equation 11. Standard errors are in
parentheses.
**
Significantly different from zero, p < 0.01.
*
Significantly different from zero, p < 0.05.

The findings reveal that the additional advertising received by the Treatment sample led
to a significant short-run increase in demand for both the Best and Other customers. The
demand increase was approximately 5.1% for the Best customers and 13.8% for the
Other customers. In percentage terms, the demand increase was significantly larger
among the Other customers, but this was calculated over a small base. In absolute terms
the effect was not significantly different across the two populations. We conclude that
current advertising can lead to a significant increase in short-run demand.

Does Current Advertising Cause Future Sales? Page 18
While these results are reassuring, they are not the main focus of this paper. Instead, we
are interested in learning how increasing current advertising affects demand in future
periods.
4.2 Does Current Advertising Impact Future Demand?
We report the impact of the Treatment on Posttest demand in Table 5. For the sake of
brevity we restrict attention to the multivariate analysis and only report the coefficients
for the Treatment variable. Complete findings are reported in Table A5 in the Appendix.
As a basis of comparison we repeat the corresponding coefficients for the Test period and
also report the coefficients when combining the data from both the Test and Posttest
periods (we label this the Total period).
Table 5: Comparison of Test Period, Posttest Period and Total Results

Other Customers Best Customers
Test Period
0.138
**
(0.019)
0.051
*
(0.010)
Posttest Period
0.097
**
(0.026)
-0.037
**
(0.013)
Total: Test and Posttest Periods
0.124
**
(0.015)
0.016
*
(0.008)
Sample Sizes
9,548
9,825
The table reports the Treatment variable coefficients when estimating
Equation 11 separately on the Test period, Posttest period and Total period
datasets. Complete findings (including the omitted coefficients) are reported
in Table A5 in the Appendix. Standard errors are in parentheses.
**
Significantly different from zero, p < 0.01.
*
Significantly different from
zero, p < 0.05.

The findings reveal a strikingly different picture for the Best and Other customers.
Amongst the Other customers the demand expansion during the Test period persists
throughout the Posttest period. The effect size drops from 13.8% in the Test period to
9.7% in the Posttest period, but remains significantly different from zero. Amongst the
Best customers we also see a significant long-run effect, however, the sign of the effect is
reversed. The increase in demand during the Test period in the Treatment condition is
offset by a significant reduction in Posttest demand. This finding for the Best customers
is consistent with temporal substitution. The increase in demand during the Test period
Does Current Advertising Cause Future Sales? Page 19
appeared to result at least in part from substitution, with customers shifting purchases
from the Posttest period to the Test period. While similar patterns of results have been
reported for price promotions, to our knowledge this is the first evidence of a significant
negative long-run effect attributed to advertising. We note that the findings cannot be
attributed to price differences as we only manipulated the frequency with which catalogs
were mailed, and not the content of the catalogs.

Recall that the Posttest period extended from January 1, 2003 through August 13, 2003.
It is possible that the adverse outcome persists beyond this period. To investigate this
possibility, we divided the Posttest period into two equal-sized (112-day) sub-periods and
repeated the analysis. This allows us to compare the impact of the additional catalog
advertising on demand at the start and end of the Posttest period. The findings for both
sub-periods are summarized in Table 6 (detailed findings are presented in Table A6 in the
Appendix).
Table 6: Comparison of Posttest Results
Start and End of the Posttest Period

Other Customers Best Customers
Start of Posttest Period
0.122
**
(0.037)
-0.094
**
(0.019)
End of Posttest Period
0.073
*
(0.037)
0.016

(0.018)
Complete Posttest Period
0.097
**
(0.026)
-0.037
**
(0.013)
Sample Sizes
9,548
9,825
The table reports the Treatment variable coefficients when estimating
Equation 11 using data from the start and end of the Posttest period.
Complete findings (including the omitted coefficients) are reported in Table
A6 in the Appendix. Standard errors are in parentheses.
**
Significantly
different from zero, p < 0.01.
*
Significantly different from zero, p < 0.05.

The negative Posttest outcome for the Best customers is concentrated at the start of the
period. By the end of the period the effect is no longer apparent. This is consistent with
our interpretation that the adverse long-run outcome for these customers reflects inter-
temporal substitution. In studies of inter-temporal substitution in the pricing literature we
Does Current Advertising Cause Future Sales? Page 20
see a similar pattern, with the post-promotion dip concentrated immediately after the
promotion period, and no effect observed on demand in later periods.

For the Other customers, the increase in catalog frequency in the Treatment condition
leads to a significant increase in demand throughout the Posttest period. Although the
estimated effect-size drops from 12.2% to 7.3% by the end of the period, the difference
between the two coefficients is not statistically significant. These findings suggest that
the favorable lift in demand for the Other customers may also have extended beyond the
Posttest Period, so that coefficient reported in Table 5 for the Total period may
underestimate the true size of the effect.

The findings in Tables 5 and 6 also reveal how the findings change as we vary the length
of the Test and Posttest periods. When the demarcation date distinguishing the Test and
Posttest periods is extended beyond December 31, 2002 to also include the start of 2003,
we see a drop in the Test Period effect among the Best customers. The Treatment effect
is most negative for these customers in the first months of 2003, and so extending the
demarcation date into 2003 leads to the inclusion of this negative long-run effect into the
Test period results. For the Other customers varying the demarcation date has little
impact on the findings.

In Section 1, we argued that the substitution interpretation for the negative long-run effect
observed among the Best customers also has a cross-channel analogy. An implication is
that if our interpretation of the results is correct, we should observe a similar effect across
channels. Mail and telephone are the primary ordering channels for catalog orders, while
alternative channels include both traditional retail stores and Internet stores. Recall that
we received demand data for purchases made through both the catalog (mail and
telephone) and the companys Internet website. In the findings reported above we
aggregated Test period demand across the catalog and Internet channels. However, by
analyzing demand separately for these two channels we can investigate whether the
incremental catalog in the treatment condition led to substitution from the Internet to the
catalog channel.
Does Current Advertising Cause Future Sales? Page 21
4.3 Cross-Channel Substitution
To distinguish the impact of the advertising manipulation on the two ordering channels
we separately calculated the number of items purchased during the Test period through
the Internet and catalog channels (our data does not distinguish between catalog orders
received via mail vs. telephone). We then re-estimated Equation 11 separately using both
of these dependent measures. The findings are reported in Table 7. Again, for ease of
presentation we only report the Treatment coefficients (the complete model is reported in
Table A7 in the Appendix).
Table 7: Comparison of Test Period Results By Channel

Other Customers Best Customers
Catalog Channel
0.116
**
(0.020)
0.063
**
(0.011)
Internet Channel
0.303
**
(0.055)
-0.096
*
(0.038)
Both Channels
0.138
**
(0.019)
0.051
**
(0.010)
Sample Sizes
9,548
9,825
The table reports the Treatment variable coefficients when estimating
Equation 11 separately on demand from the catalog channel, demand from
the Internet channel, and total demand across both channels. Complete
findings (including the omitted coefficients) are reported in Table A7 in the
Appendix. Standard errors are in parentheses.
**
Significantly different
from zero, p < 0.01.
*
Significantly different from zero, p < 0.05.

The pattern of findings in the cross-channel analysis is analogous to the long-run
analysis. The favorable outcome for Other customers extends across both channels. In
contrast, among Best customers the favorable outcome in the Catalog channel is offset by
a significant reduction in demand over the Internet channel. We caution that we do not
have data describing demand in the companys retail stores. The evidence of channel-
switching among the Best customers suggests that the increase in catalog advertising may
also have switched demand from the retail stores to the catalog channel, at least for
customers living close to these stores. In this respect, our measure of the total change in
Test period demand (in Table 6) may understate the effect; that is, it could be positively
biased for the Best customers and negatively biased for the Other customers.

Does Current Advertising Cause Future Sales? Page 22
4.4 Sending Catalogs to Their Best Customers
As we discussed, most companies design their catalog mailing policies by varying their
mailing policies and evaluating only the orders received from those same catalogs. This
myopic focus on the short-run catalog demand ignores the externalities in other channels
and in future periods. For example, the findings in Table 5 indicate that among Best
customers the short-run response to advertising overstates the long-run response to
advertising by a factor of three (5.1% vs. 1.6%). As a result, firms that rely on the short-
run response are likely to overinvest in advertising.

To illustrate the implications of this on firm profit, we summarize the profits earned in
each condition in Table 8. The profits are calculated as the sum of the items ordered by
each customer, multiplied by the profit margin on each item, less catalog printing and
mailing costs incurred during the Test period. We compare three different profit
measures: (1) profit earned from the catalog channel in the Test period; (2) profit earned
from all channels in the Test period (including Internet orders); and (3) profit earned from
all channels in both the Test and Posttest periods.

Focusing first on the Best customers, we see that if the company focused solely on profits
earned during the Test period from the catalog channel it would erroneously conclude
that it is profitable to send catalogs more frequently to its Best customers. After allowing
for the adverse inter-temporal and cross-channel outcomes we see that the profit result is
reversed. The company actually earned a higher average profit in the Control condition.
Among the Other customers the positive externalities in the Internet channel and Posttest
period almost lead to the opposite outcome. Mailing more frequently to the Other
customers is clearly more profitable when these externalities are taken into account.
However, this conclusion is much weaker if attention is restricted to Test period profits
from the Catalog channel.

Does Current Advertising Cause Future Sales? Page 23
Table 8: Average Profit Earned Per Customer
Treatment vs. Control
Control Treatment Difference
Best Customers
Catalog profit during test period $89.98 $91.56
$1.58
Total test period (including Internet) $98.74 $100.27
$1.53
Total profit (including Posttest and Internet) $164.57 $163.84
-$0.73
Sample size 4,921 4,904

Other Customers
Catalog profit during test period $15.50 $15.86
$0.36
Total test period (including Internet) $19.46 $20.54
$1.08
Total profit (including Posttest and Internet) $35.06 $37.49
$2.43
Sample size 4,790 4,758

Profits earned from each customer are calculated as the sum of the items ordered by each
customer, multiplied by the profit margin on each item, minus the cost of printing and mailing
catalogs during the Test period.

This interpretation of the findings raises the question as to why companies ignore these
long-run and cross-channel effects. We offer to two responses. First, not all catalog
firms have ignored these effects. For example, Rhenania, a German book catalog
company, revised its mailing policies to shift its objective function from maximizing
short-run profits to also consider profits in future periods (Elsner, Krafft and
Huchzermeier 2003). The company attributed the reversal of its history of declining
sales, market share, and profits to the adoption of its new mailing policy.

Our second response is that measuring and responding to long-run and cross-channel
effects are difficult. Consider first the measurement problem. When customers call to
place an order over the telephone they are asked for a code printed on the catalog that
identifies which catalog customers are ordering from. Similarly, when a customer orders
via mail using the form bound into a catalog, companies can again identify the catalog
from a code pre-printed on the order form. As a result, companies can construct a rich
database identifying which of the customers who received a catalog placed an order
through the catalog channel. In contrast, when a customer places an order through a
companys Internet website, it is generally not possible to identify whether the order was
Does Current Advertising Cause Future Sales? Page 24
prompted by a catalog, and (if so) which catalog the customer is ordering from. Linking
future purchases to past mailing decisions is even more difficult.

Furthermore, when future purchases are linked to past mailings as part of a controlled
experiment, it turns out to be important to consider different customer subsets separately.
If the Best and Other customers are pooled, then the net effect of additional advertising
on future sales is statistically indistinguishable from zero. This is not because the effect
on individual consumers is zero, however. Instead, it reflects the negative effects on the
Best customers canceling out the continuing positive contributions for the Other
customers. This could have been easily overlooked when analyzing historical data (in the
absence of a controlled experiment).

Even when companies can effectively measure cross-channel and long-run customer
response functions, optimizing the companys mailing strategy remains difficult.
Optimizing the short-run policy is relatively straight-forward as there are only two
possible actions: mail or dont mail. In contrast, the long-run mailing policy has an
infinite range of possible mailing sequences. Moreover, evaluating the profitability of
these sequences is no long a straight-forward statistical problem. Some catalog
companies have tested sequences of mailing policies using split-sample field tests. Yet
such approaches cannot reveal the optimal policy without requiring an infinite series of
such tests, while evaluating the long-run impact of these tests requires that companies
wait for the long-run to occur.

At least one important question remains. In our model of the relationship between
advertising and future demand we assumed that the relationship between advertising and
preferences is positive but did not offer an explanation for this relationship. The
literature suggests two competing explanations: information and persuasion. Our ability
to compare how the long-run outcome varied across different samples of customers
provides an opportunity to distinguish between these two explanations.

Does Current Advertising Cause Future Sales? Page 25
4.5 Information or Persuasion?
Recall that the catalogs used in this study contain information about what products are
available and the current prices of those products. Yet prices and product availability
change quickly in the womens clothing category, and so the incremental information that
customers in the Treatment condition received about product availability and prices is
unlikely to have significantly affected their future demand. With few exceptions, this
information would no longer have been relevant once the Posttest period started (almost
four months after the last manipulation).

It is tempting to conclude that the incremental advertising sent to customers in the
Treatment condition therefore served a persuasive role. However, before doing so, it is
important to also investigate whether the incremental catalogs may have provided other
types of information (other than prices and product availability). Recall that the catalogs
used in this study also contain descriptions of the companys warranties, ordering
procedures, and payment methods, together with implicit cues and explicit claims that
may provide customers with information about product quality. It is possible that by
exposing customers to this static information, the additional catalog advertising served an
informational role that may have influenced future demand.

A recent study by Ackerberg (2001) suggests an approach for distinguishing between
these informative and persuasive interpretations. In a study of supermarket demand for
yogurt, Ackerberg argues that the informative role of advertising should only affect the
behavior of customers who have a need for information. In particular, the effect should
be limited to customers who have relatively little experience with the company. Using
historical data from a panel of households, he shows that inexperienced customers
demonstrate a stronger reaction to television advertising than experienced customers, and
concludes that this is evidence that the advertising is performing an informative role. We
can use a similar approach to distinguish between the persuasive and informative
explanations for the long-run findings in this paper.

Does Current Advertising Cause Future Sales? Page 26
Ackerberg argued that the informative role of advertising should only affect the behavior
of customers who have a need for information. There should be an analogous moderating
effect for the persuasive role of advertising: customers who are all already convinced
about the merits of the company should not be susceptible to additional persuasion. This
reasoning leads to the two-dimensional interaction summarized in Table 9, predicting
which customers are susceptible to additional information and/or persuasion.
Table 9: Identifying Which Customers Are Susceptible
To Additional Information and/or Persuasion

Persuaded
Customers
Unpersuaded
Customers
Informed Customers Persuasive Role
Uninformed Customers Informative Role
Informative Role
and Persuasive Role

Operationalizing this interaction requires that we identify a measure to distinguish
customers who are informed from those who are uninformed. As we discussed, the
information at issue is unlikely to be information about current prices and/or product
availability. Instead, the relevant information is more likely to concern static
characteristics of the company and its products and policies. The experimental
manipulation involved mailing additional catalogs and so a good measure of how much
prior information customers had about these static characteristics is how many catalogs
the customers had received prior to the start of the study. Customers who had already
received a lot of catalogs should already have a lot of prior information about the
company. If the catalog advertising acts to inform customers about the static
characteristics of this company we should not see a response from these customers.
Therefore, we use the total number of catalogs received in the five years prior to the start
of the experimental manipulation (Catalogs Received) as a measure of the degree to
which customers are informed.

We also need a measure to distinguish customers who at the time of the test are already
persuaded about the merits of the company from those who are not yet persuaded.
Does Current Advertising Cause Future Sales? Page 27
Customers who are convinced about the merits of the company are more likely to have
recently purchased a relatively large number of items. Therefore, as a measure of prior
persuasion we use the total expenditure in the 24-months prior to the start of the
experimental manipulation (Prior Expenditure).

Before presenting any analysis it is appropriate to consider the sources of variation in the
prior mailing policies and customers prior purchasing levels. Our approach for
distinguishing between the persuasion and information explanations requires that there is
independent variation between the Catalogs Received and Prior Expenditure measures.
The correlation between the two measures is 0.325 (significantly different from zero,
p<0.01). This correlation is consistent with our analysis of the impact of catalog
advertising on short-run demand: customers who have received more catalogs are more
likely to have purchased (see Table 4). However, further investigation reveals that there
are two segments of customers who lower the correlation and contribute to independent
variation in the measures. We can illustrate these two segments by using the medians of
the Catalogs Received and Prior Expenditure measures to split customers into four
segments (see Table 10).

Approximately a third of the customers have high prior expenditure and have received
many catalogs, while another third have relatively low prior expenditure and have
received fewer catalogs. The large number of customers in these two segments is
consistent with a causal relationship between historical mailing decisions and historical
demand. Yet we also see approximately 3,000 customers in each of the off-diagonal
cells. The 3,056 customers with high expenditure in the previous two years but few
catalogs received are typically customers whose first purchase from the company was
relatively recent. A recent first purchase means that the company has had relatively few
opportunities to send catalogs to this customer. Although the volume of their recent
purchases indicates that they have favorable perceptions of the company, the customers
have received relatively little information from the companys catalogs.

Does Current Advertising Cause Future Sales? Page 28
Table 10: Samples Sizes that Result From Median Splits of the Catalogs Received
and Prior Expenditure Measures
High Prior
Expenditure
Low Prior
Expenditure
Many Catalogs Received 6,739 3,050
Few Catalogs Received 3,056 6,528
The table reports the sample sizes that result when splitting the sample into sub-
samples using a median split of both the Prior Expenditure and Catalogs Received
variables.

The 3,050 customers who have received many prior catalogs but have made few recent
purchases are generally customers whose first purchase was made well before the start of
the manipulation period. Recall that the Catalogs Received measure considers all
catalogs mailed in the five years prior to the manipulation period, while the Prior
Expenditure measure considers purchases within two years of the manipulations.
Customers who purchased between two and five years before the test will have received
many catalogs but may have made few recent purchases. Although these customers will
be well-informed about the static characteristics of the company, their lack of recent
purchases suggests that there is an opportunity to raise their persuasion levels.

To estimate the moderating role played by prior information and prior persuasion we
modified Equation 11 to incorporate interactions between the Treatment variable and
both Catalogs Received and Prior Expenditure:

0 1 2 3
4 5 6
7 8
log( ) log( ) log( )
i i i
i i i
i i i
Recency Frequency Monetary Value
Catalogs Received Prior Expenditure Treatment
Treatment * Catalogs Received Treatment * Prior Expenditure
= + + +
+ + +
+ +
X
i
i
(12)

Under this specification the
7
and
8
coefficients estimate the moderating influence of
the two measures on the long-run impact of the treatment. We report these findings in
Table 11, where we also report three benchmark models that include each of these
interactions separately and neither of the interactions. In each of the models we pool data
Does Current Advertising Cause Future Sales? Page 29
from both the Best and Other customer samples. We again remind readers that this
analysis preserves the experimental control due to the random assignment of customers to
the two experimental conditions.
Table 11: The Impact of Additional Catalog Advertising on Posttest Demand The
Moderating Role of Catalogs Received and Prior Expenditure
Model 1 Model 2 Model 3 Model 4
Intercept
-1.747
**
(0.114)
-1.789
**
(0.115)
-1.731
**
(0.115)
-1.749
**
(0.116)
Recency
-0.183
**
(0.004)
-0.183
**
(0.004)
-0.183
**
(0.004)
-0.183
**
(0.004)
Frequency
0.617
**
(0.010)
0.617
**
(0.010)
0.602
**
(0.010)
0.602
**
(0.011)
Monetary Value
0.266
**
(0.025)
0.264
**
(0.025)
0.257
**
(0.025)
0.257
**
(0.025)
Catalogs Received
-0.031
*
(0.014)
-0.002

(0.018)
-0.026

(0.014)
-0.013

(0.018)
Prior Expenditure
0.010
**
(0.0004)
0.011
**
(0.0004)
0.015
**
(0.0006)
0.015
**
(0.0006)
Treatment
-0.025
*
(0.012)
0.075
*
(0.037)
0.060
**
(0.014)
0.100
**
(0.037)
Treatment * Catalogs Received

-0.058
**
(0.020)

-0.024

(0.020)
Treatment * Prior Expenditure

-0.007
**
(0.0007)
-0.007
**
(0.0007)
Log Likelihood -42,046 -42,042 -41,993 -41,992
Sample Size 19,373 19,373 19,373 19,373
The table reports the coefficient estimates that result from estimating Equation 12 on Posttest demand.
Standard errors are in parentheses.
**
Significantly different from zero, p < 0.01.
*
Significantly
different from zero, p < 0.05.

There are several findings of interest. First, the interaction between the Treatment effect
and Prior Expenditure (in Models 3 and 4) is negative and highly significant. This
indicates that the long-run impact of the experimental manipulation was moderated by the
level of customers recent prior expenditure. The favorable long-run outcome was
limited to customers with relatively low levels of prior expenditure; a finding which is
consistent with the earlier contrasting results for the Best and Other customer samples.

Does Current Advertising Cause Future Sales? Page 30
Second, the interaction between the Treatment effect and Catalogs Received is also
negative. However, this coefficient is only significant in Model 2, where the interaction
with Prior Expenditure is omitted. It is possible that Prior Expenditure acts as a proxy
for Catalogs Received in Model 2. When both interaction terms are included (Model 4)
the interaction between the Treatment and Catalogs Received is no longer significant. A
likelihood ratio test between Models 3 and 4 indicates that the addition of the Catalogs
Received interaction term does not increase the explanatory power of the model. We
conclude that the findings suggest that the number of prior Catalogs Received does not
independently moderate the treatment effect.

Finally, in the absence of both interaction terms (Model 1) we see that the coefficient for
the Treatment variable is negative and significant. This contrasts with the findings in
Models 2, 3 and 4, and highlights the difficulty of analyzing the long-run response to
advertising. Failure to anticipate the heterogeneity in the long-run response would have
led to the (erroneous) conclusion that the intervention had a negative long-run impact on
all customers.

We conclude that the findings in Table 11 offer support for the persuasive view. The
expansion in long-run demand does not appear to apply to the companys most valuable
customers. Intuitively, it is hard to make the companys best customers any better as
these customers are already using the company to satisfy their category needs - there is a
limit to how many clothes even the most loyal customer can wear!

5. Conclusions
We have reported the findings from a large-scale field study in which we exogenously
manipulated the frequency of catalog advertising sent to randomly selected customer
samples. We then tracked both the immediate response and the impact on future
purchases by these customers. The findings confirm that current advertising can impact
future demand. Interestingly, the impact is quite heterogeneous. Among the companys
most valuable customers, who had purchased recently and frequently from the company,
Does Current Advertising Cause Future Sales? Page 31
the long-run impact was negative. The short-run lift in demand for these customers was
apparently largely due to cross-channel and temporal substitution.

In contrast, among the less valuable customers, who had purchased less frequently and/or
less recently, advertising had a positive impact on future demand. We note that these
customers are also susceptible to temporal substitution, and so the favorable long-run
outcome occurred despite the adverse effects of temporal substitution (if any).

We investigated two alternative explanations for this long-run outcome. The first
explanation focused on the persuasive role of advertising and the second focused on
advertisings informative role. To distinguish these explanations we took advantage of
predicted heterogeneity in how sensitive different types of customers are to additional
information and/or additional persuasion. The findings offer support for the persuasive
argument. Customers most affected by the advertising were those whose past behavior
indicated that there was an opportunity to increase their consumption through additional
persuasion.

The findings also offer an explanation for a question that has often left customers
perplexed: why do companies send so many catalogs to their best customers? It seems
that the intensive mailing frequency to a companys best customers can be explained in
part by a (mistaken) focus on short-run outcomes when designing catalog mailing
policies. If a company overlooks the negative externalities on future demand and demand
in other channels, it will tend to over-mail to its best customers. The same myopic focus
may lead to the opposite outcome for other less valuable customers. For these
customers the externalities are positive, so that it may be profitable to mail to customers
who are unlikely to purchase immediately, as by doing so companies can increase the
probability of a future purchase.

Our findings help untangle the questions about advertisings effects on long run demand.
It turns out that advertising causes both increases and decreases in future demand,
depending on the type of customer. We can identify which types of customers are likely
Does Current Advertising Cause Future Sales? Page 32
to be in each group. Our results also demonstrate the power of field experiments, not
only for advancing research on the economics of advertising, but also in identifying
potential gaps in business practice.
Does Current Advertising Cause Future Sales? Page 33
6. References
Aaker D. and J. M. Carman (1982), Are Your Overadvertising? Journal of Advertising,
22(4), 57-70.
Ackerberg, D. A. (2001), Empirically Distinguishing Informative and Prestige Effects of
Advertising, Rand Journal of Economics, 32(2), pp 316-33.
Ackerberg, D. A. (2003) Advertising, Learning, and Consumer Choice in Experience
Good Markets: A Structural Empirical Examination, International Economic
Review, 44(3), August, 1007-40.
Advertising Age (2005), Fact Pack: 3
rd
Annual Guide to Advertising and Marketing,
February 28, Crain Communications, Inc.
Anderson E. T. and D. I. Simester (2004), Impact of Promotion Depth on New vs.
Established Customers, Marketing Science, 23(1), 4-20.
Bagwell, K. (2005), The Economic Analysis of Advertising, forthcoming in Mark
Armstrong and Rob Porter (eds.), Handbook of Industrial Organization, 3, North-
Holland: Amsterdam.
Becker G. S. and K. M. Murphy (1993), A Simple Theory of Advertising as a Good or
Bad, Quarterly Journal of Economics, 108, 942-964.
Benham, L. (1972), The Effects of Advertising on the Price of Eyeglasses, Journal of
Law and Economics, 15, 337-352.
Blattberg, R. and S. Neslin (1990), Sales Promotions, Prentice Hall.
Borden, N. H. (1942), The Economic Effects of Advertising, Richard D. Irwin.
Comanor, W. S. and T. A. Wilson (1967), Advertising, Market Structure and
Performance, The Review of Economics and Statistics, 49, pp. 423-40.
Comanor, W. S. and T. A. Wilson (1974), Advertising and Market Power, Harvard
University Press.
Direct Marketing Association (2003), The DMA 2003 Statistical Fact Book, New Jersey:
DMA.
Direct Marketing Association (2005), The DMA 2005 Statistical Fact Book, New Jersey:
DMA.
Elsner R, M. Krafft, and A. Huchzermeier (2003). Optimizing Rhenanias Mail-order
Business through Dynamic Multilevel Modeling, Interfaces, Vol. 33(1), 50-66.
Erdem T. and M. Keane (1996), Decision-Making Under Uncertainty: Capturing
Dynamic Brand Choice Processes in Turbulent Consumer Goods Markets,
Marketing Science, 15(1), 1-20.
Gnl, F. and M. Shi (1998). Optimal Mailing of Catalogs: A New Methodology Using
Estimable Structural Dynamic Programming Models, Management Science,
44(9) 1249-62.
Does Current Advertising Cause Future Sales? Page 34
Grossman, G. and C. Shapiro (1984), Informative Advertising with Differentiated
Products, Review of Economic Studies, 51, 63-81.
Hendel, I. and A. Nevo (2002), Sales and Consumer Inventory, NBER working paper
9048, NBER Cambridge MA.
Hendel, I. and A. Nevo (2003), The Post Promotion Dip Puzzle: What do the Data Have
to Say? Quantitative Marketing and Economics, 1(4), 409-424.
Hendel, I. and A. Nevo (2005), Measuring the Implications of Sales and Consumer
Inventory Behavior, NBER working paper 11307, NBER Cambridge MA.
Ippolito P. M. and A. D. Mathios (1990), Information, Advertising and Health Choices:
A Study of the Cereal Market, RAND Journal of Economics, 21, 459-480.
Kaldor N. (1950), The Economic Aspect of Advertising, Review of Economic Studies
18 (1), 1-27.
Kihlstrom R. and M. Riordan (1984), Advertising as a Signal, Journal of Political
Economy, 92, 427-450.
Krishnamurthi, L. and S. P. Raj (1985), The Effect of Advertising on Consumer Price
Sensitivity, Journal of Marketing Research, 22, May, 119-29.
Lambin, J. J. (1976), Advertising, Competition and Market Conduct in Oligopoly Over
Time, Amsterdam: North-Holland.
Lodish, L.M., M. M. Abraham, S. Kalmenson, J. Livelsberger, B. Lubetkin, B.
Richardson, and M. E. Stevens (1995a), How T.V. Advertising Works: A Meta-
Analysis of 389 Real World Split Cable T.V. Advertising Experiments, Journal
of Marketing Research, 32(2), 125-139.
Lodish, L. M., M. M. Abraham, J. Livelsberger, B. Lubetkin, B. Richardson, and M. E.
Stevens (1995b), A Summary of Fifty-Five In-Market Experimental Estimates of
the Long-Term Effect of TV Advertising, Marketing Science, 14(3), part 2,
G133-G140.
Milgrom P. and J. Roberts (1986), Price and Advertising Signals of Product Quality,
Journal of Political Economy, 94, 796-821.
Milyo, J. and J. Waldfogel (1999), The Effect of Price Advertising on Prices: Evidence
in the Wake of 44 Liquormart, American Economic Review, 89 (5), 1081-1096.
Nelson, P. (1970), Information and Consumer Behavior, Journal of Political Economy,
78, May 311-29.
Nelson, P. (1974), Advertising as Information, Journal of Political Economy, 82,
August, 729-54.
Roberts, M. J. and L. Samuelson (1988), An Empirical Analysis of Dynamic Non-Price
Competition in an Oligopolistic Industry, Rand Journal of Economics, 19, 200-
20.
Schmalensee, R. (1972), The Economics of Advertising, Amsterdam: North-Holland.
Does Current Advertising Cause Future Sales? Page 35
Schmalensee, R. (1978), A Model of Advertising and Product Quality, Journal of
Political Economy, 86 (3), 485-503.
Seldon B. J. and K. Doroodian (1989), A Simultaneous Model of Cigarette Advertising:
Effects on Demand and Industry Response to Public Policy, The Review of
Economics and Statistics, 71, 673-77.
Simester, D. I., Sun P. and J. N. Tsitsiklis (2005), Optimal Catalog Mailing Policies,
Management Science, forthcoming.
Stigler, G. J. (1961), The Economics of Information, Journal of Political Economy, 69,
June, 213-25.
Telser, L. G. (1964), Advertising and Competition, Journal of Political Economy,
72(6), December, 537-62.

Does Current Advertising Cause Future Sales? Page 36
Table A5: Comparison of Test Period, Posttest Period and Total Results
Posttest Period Total Period

Other
Customers
Best
Customers
Other
Customers
Best
Customers
Intercept
-0.872
**
(0.168)
-3.195
**
(0.146)
-0.432

(0.105)
-3.168
**
(0.092)
Recency
-0.289
**
(0.008)
-0.146
**
(0.005)
-0.281
**
(0.005)
-0.137
**
(0.003)
Frequency
0.465
**
(0.013)
0.723
**
(0.010)
0.480
**
(0.008)
0.738
**
(0.006)
Monetary Value
0.221
**
(0.037)
0.515
**
(0.032)
0.350
**
(0.023)
0.705
**
(0.020)
Treatment
0.097
**
(0.026)
-0.037
**
(0.013)
0.124
**
(0.015)
0.016
*
(0.008)
Log Likelihood
-13,145
-28,891
-25,422
-44,298
Sample Size
9,704
9,834
9,704
9,834
The Posttest findings reports the coefficients from Equation 11 estimated using data from the
Posttest period. The Total Period findings report the coefficients from Equation 11 estimated using
data from the entire period (Test and Posttest). Standard errors are in parentheses.
**
Significantly different from zero, p < 0.01.
*
Significantly different from zero, p < 0.05.

Does Current Advertising Cause Future Sales? Page 37
Table A6: Comparison of Posttest Results
Start and End of the Posttest Period
Start of Posttest Period End of Posttest Period

Other
Customers
Best
Customers
Other
Customers
Best
Customers
Intercept
-2.200
**
(0.256)
-4.656
**
(0.209)
-1.033
**
(0.217)
-3.172
**
(0.202)
Recency
-0.272
**
(0.012)
-0.162
**
(0.008)
-0.305
**
(0.011)
-0.131
**
(0.007)
Frequency
0.466
**
(0.019)
0.754
**
(0.014)
0.467
**
(0.019)
0.692
**
(0.013)
Monetary Value
0.345
**
(0.056)
0.688
**
(0.047)
0.117
**
(0.048)
0.354
**
(0.045)
Treatment
0.122
**
(0.037)
-0.094
**
(0.019)
0.073
*
(0.037)
0.016

(0.018)
Log Likelihood
-8,092
-19,387
-8,254
-20,447
Sample Size
9,704
9,834
9,704
9,834
The Internet Channel findings reports the coefficients from Equation 11 estimated using purchases
from the start and end of the Posttest period. Standard errors are in parentheses.
**
Significantly different from zero, p < 0.01.
*
Significantly different from zero, p < 0.05.
Does Current Advertising Cause Future Sales? Page 38
Table A7: Comparison of Test Period Results By Channel
Internet Channel Catalog Channel

Other
Customers
Best
Customers
Other
Customers
Best
Customers
Intercept
-2.460
**
(0.343)
-9.563
**
(0.422)
-1.491
**
(0.144)
-4.110
**
(0.122)
Recency
-0.447
**
(0.016)
-0.066
**
(0.016)
-0.249
**
(0.007)
-0.136
**
(0.004)
Frequency
0.598
**
(0.028)
0.829
**
(0.028)
0.473
**
(0.011)
0.742
**
(0.008)
Monetary Value
0.326
**
(0.075)
1.362
**
(0.093)
0.438
**
(0.031)
0.783
**
(0.027)
Treatment
0.303
**
(0.055)
-0.096
*
(0.038)
0.116
**
(0.020)
0.063
**
(0.011)
Log Likelihood
-4,739
-9,623
-17,882
-32,568
Sample Size
9,704
9,834
9,704
9,834
The Internet Channel findings reports the coefficients from Equation 11 estimated using purchases
through the Internet channel in the Test period. The Catalog Channel findings use purchases through
the Catalog channel.

**
Significantly different from zero, p < 0.01.
*
Significantly different from zero, p < 0.05.